Reference · Usage

How usage is counted

Every call spends credits, and calls cost different amounts because they are different work. A prior-art search runs a rerank over the corpus; reading an examiner's citations is one indexed lookup. Charging both the same would mean either pricing the cheap calls for the expensive one, or letting a client loop the expensive one at the cheap price.

The same tariff applies whether you call the REST endpoint or the MCP tool in front of it, because the tool proxies to the endpoint with your own key. Plans and prices are on the pricing page.

What each call costs

Four bands, so the cost of a call is something you can hold in your head rather than a per-endpoint tariff nobody reads.

Call	What it does	Credits
Prior-art search search_prior_art	Two-stage retrieval over 10.5M patents: vector recall, then a rerank against your claim text. Every hit carries the passage it matched on.	10
Claim support mapping map_claim_support	Grounds every claim term in the specification, and for a CIP or continuation in the parent specs — which is what decides whether a term is new matter.	10
Locate in patent space locate_in_patent_space	Places an idea in the corpus and returns its neighbours, CPC sections and art units.	3
Whitespace find_whitespace	Regions of the space with thin coverage, filtered by section and opportunity.	3
Patent neighbourhood get_patent_neighborhood	What sits closest to a given filing, by embedding rather than by classification.	1
Examiner citations get_patent_citations	What an examiner actually cited against an application, from 239M citation edges.	1
Antecedent-basis check check_antecedent_basis	§112(b) antecedent-basis defects in a draft or filed claim set. Deterministic, no model call.	1

Drawings cost 2. Map tiles, manifests, and anything not listed cost 1. An endpoint that appears without a price in this table costs 1, which is the safe direction: a new cheap route billed at 1 is right, and a new expensive one shows up as our problem rather than as your outage.

What each plan covers

The same allowance read two ways. The left column is what you plan against; the right is what the meter counts.

Plan	Covers	Credits / mo
Evaluate	About 50 legal context searches a month	500
Launch	Roughly 250 searches a month, and headroom past it	2,500

The "covers" figure assumes every call is a prior-art search, which is the most expensive thing you can do. A realistic mix of searches, citation lookups and neighbourhood calls goes considerably further.

Ceilings and headroom

There are three ceilings, and they do different jobs. Two of them you will never notice.

Monthly credits: The allowance your plan includes, reset on the first of the month. On a paid plan this is a soft line: calls keep being served past it, nothing is charged automatically, and we come and talk to you about sizing. On the free plan it is hard, because there it is an abuse control rather than a billing one.
Purchased credits: Credits bought as a top-up, or committed under an agreement. They do not reset and do not expire, and they are only spent once the monthly allowance is gone — so covering a busy month in advance is not a bet on that month.
Requests per day: A burst guard, not a product limit. It exists so a runaway loop cannot spend a month in an hour, which protects your allowance at least as much as our capacity. Normal use does not reach it.

Reading it from code

You should not have to poll a dashboard to know where you stand.

X-Antecedint-Plan-Overage: true

Present when the call was served from the headroom past your monthly allowance. It is not an error and the response is normal — treat it as the cue to look at your plan size, not as something to retry or back off from.

429 · USAGE_LIMIT_EXCEEDED

A ceiling was actually reached. The body reports the daily and monthly figures the decision used, and credits.required says what the refused call would have cost, so you can tell a large call hitting a small remainder from an allowance genuinely spent.

503 · FREE_TIER_CAPACITY

Free-tier evaluation capacity is momentarily full. Nothing to do with your allowance — it is a shared pool, and Retry-After says when to come back. Paid plans are not subject to it.

Throughput

Separate from credits, and the ceiling a production workload usually meets first. Rate limits are a property of your key rather than of the deployment, so they move with your plan.

Plan	Throughput
Evaluate	60 messages/min · 15 tool calls/min
Launch	300 messages/min · 60 tool calls/min
Enterprise	Committed, and dedicated rerank capacity

A JSON-RPC batch counts as one request for rate limiting but spends the credits of every tool call inside it, so batching is a way to save round trips rather than credits.